Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Privacy Meets Explainability: Managing Confidential Data and Transparency Policies in LLM-Empowered Science

Yashothara Shanmugarasa, Shidong Pan, Ming Ding, Dehai Zhao, Thierry Rakotoarivelo

TL;DR

This paper tackles confidentiality and explainability in LLM-powered scientific workflows by introducing DataShield, a three-module framework comprising confidential data detection, policy summarization (including external tool policies and internal codes of conduct), and a privacy-focused visualization dashboard. The approach combines rule-based and retrieval-augmented detection for confidential data, a two-layer policy summarization (using PoliGraph and LLM-RAG) to generate concise privacy nutrient labels, and an integrated dashboard that presents data flow, tool policies, and compliance. Early results compare confidential-data detection across baselines and local models, showing trade-offs in accuracy and precision, while outlining a rigorous user-study plan to evaluate usability, trust, and effectiveness. The work aims to empower scientists to recognize data handling risks, ensure policy alignment, and foster trustworthy adoption of LLMs in research workflows, with potential extensions to broader domains and collaborative environments.

Abstract

As Large Language Models (LLMs) become integral to scientific workflows, concerns over the confidentiality and ethical handling of confidential data have emerged. This paper explores data exposure risks through LLM-powered scientific tools, which can inadvertently leak confidential information, including intellectual property and proprietary data, from scientists' perspectives. We propose "DataShield", a framework designed to detect confidential data leaks, summarize privacy policies, and visualize data flow, ensuring alignment with organizational policies and procedures. Our approach aims to inform scientists about data handling practices, enabling them to make informed decisions and protect sensitive information. Ongoing user studies with scientists are underway to evaluate the framework's usability, trustworthiness, and effectiveness in tackling real-world privacy challenges.

Privacy Meets Explainability: Managing Confidential Data and Transparency Policies in LLM-Empowered Science

TL;DR

This paper tackles confidentiality and explainability in LLM-powered scientific workflows by introducing DataShield, a three-module framework comprising confidential data detection, policy summarization (including external tool policies and internal codes of conduct), and a privacy-focused visualization dashboard. The approach combines rule-based and retrieval-augmented detection for confidential data, a two-layer policy summarization (using PoliGraph and LLM-RAG) to generate concise privacy nutrient labels, and an integrated dashboard that presents data flow, tool policies, and compliance. Early results compare confidential-data detection across baselines and local models, showing trade-offs in accuracy and precision, while outlining a rigorous user-study plan to evaluate usability, trust, and effectiveness. The work aims to empower scientists to recognize data handling risks, ensure policy alignment, and foster trustworthy adoption of LLMs in research workflows, with potential extensions to broader domains and collaborative environments.

Abstract

As Large Language Models (LLMs) become integral to scientific workflows, concerns over the confidentiality and ethical handling of confidential data have emerged. This paper explores data exposure risks through LLM-powered scientific tools, which can inadvertently leak confidential information, including intellectual property and proprietary data, from scientists' perspectives. We propose "DataShield", a framework designed to detect confidential data leaks, summarize privacy policies, and visualize data flow, ensuring alignment with organizational policies and procedures. Our approach aims to inform scientists about data handling practices, enabling them to make informed decisions and protect sensitive information. Ongoing user studies with scientists are underway to evaluate the framework's usability, trustworthiness, and effectiveness in tackling real-world privacy challenges.

Paper Structure

This paper contains 15 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Research Questions
  3. Related Works
  4. Studies on Personally Identifiable Information Detection
  5. Studies on Policy Summarization and Privacy Nutrition Labels
  6. Studies on Privacy Dashboards
  7. Studies on Explainability and Human Involvement in Responsible LLMs
  8. Our Unique Contributions
  9. Proposed solution
  10. A Confidential Data Detection Module
  11. A Policy Summarization Module
  12. A Visualization Dashboard Module
  13. User Study
  14. Preliminary Results
  15. Discussion and Future Work

Figures (2)

  • Figure 1: The overview of "DataSheild" framework with three modules output: User Interface: Clear version can be found https://drive.google.com/file/d/1FRNvTNWgI-Jnwr4Xj8Mo3KGTu-NoH7gy/view?usp=sharing
  • Figure 2: Workflow Diagram of "DataShield" – Explainable Shield for Confidential Data (RAG: Retrieval-Augmented Generation)